Peter Banks just published Debt Killed the Soviet Empire at the Boyd Institute. The argument: fiscal exhaustion under inelastic obligations dissolved the USSR through a specific sequence. Extensive-growth model exhausted by the 1970s. Oil rents collapsed 1985-86 when Brent halved. State pushed onto international debt markets it could not service in hard currency. Vneshekonombank suspended principal payments in December 1991; the Belavezha Accords dissolved the Union four days later. The piece then locates the contemporary US on the same structural axis, with finite but real buffers (Japan exception, own-currency caveat, $970B 2025 net interest already crowding out vital fiscal capacity).
The mechanism is one of the most legible failure modes in the historical record. Rome, the Ancien Régime, post-WWI Habsburgs, Argentina serially. Banks's piece is structurally responsible: qualifies the analogy, names the caveats, closes with the standalone-true observation that interest is mechanically eating fiscal capacity. The structural correction is narrow. The contemporary US is on a finite-buffer axis. It is not on the Soviet axis.
Restate Banks's mechanism structurally: a fiscal engine has inelastic obligations and a payer base. When the payer base contracts faster than obligations can be cut, the engine borrows. When borrowing service exceeds the engine's ability to raise the medium it borrowed in, the engine fails.
The Soviet engine carried three inelastic obligations summing to ~30% of GNP. Military (15-17% of GNP). Empire-subsidies (cheap oil and electricity to CMEA puppets; intra-union transfers funding Central Asia, Caucasus, Siberia from western productive provinces). Household-consumption subsidies (retail subsidies rising 4% → 20% of state budget 1965-1980s; food imports). The payer was oil rent. Brent halved 1985-86; hard-currency export earnings dropped from $32B (1983-84) to $25.1B (1986), with oil alone collapsing from $15.2B average to $7B by year-end 1986. The closed economy could not substitute another hard-currency source. The medium-mismatch was the binding step: hard-currency-denominated debt, ruble-denominated internal economy, no inflation path to wipe obligations.
The engine failed because it lost its payer.
The contemporary US fiscal engine is doing the opposite. It is acquiring a payer of unprecedented characteristics.
Tether holds $141 billion in US Treasuries as reserve composition for issued stablecoin float. That is a larger holding than most sovereigns. Q1 2026 stablecoin transactional volume hit $28 trillion, orders of magnitude larger than BTC or any prior crypto rail. The GENIUS Act, signed July 2025, requires 1:1 reserves in USD or US Treasuries for stablecoin issuance under US jurisdiction. The reserve is mechanical: per dollar of stablecoin issued under the regime, ~1 dollar of Treasury demand is generated by construction.
The AI-agent transition scales this. Per The Payer Question, most AI agents will not transact permissionlessly through Monero or CoinJoin'd BTC. They will transact through KYC-routed stablecoin rails inside corporate stacks whose regulatory posture decides the volume. Anthropic, OpenAI, Google, and the major model platforms operate inside US compliance frameworks because their business models depend on regulatory tolerance. The path of least resistance is issuer-mediated stablecoin transfer, USD-denominated, with the issuer absorbing the compliance overhead. As the AI-agent population scales geometrically through this decade, the issuer-reserve-composition channel becomes a structurally larger Treasury demand source than foreign sovereigns ever were.
Foreign-sovereign Treasury holdings have been flat at ~$8 trillion since 2014 while marketable debt tripled. The marginal Treasury buyer migrated to private accounts: money-market funds, US households, stablecoin issuers. The trajectory in that last category accelerates with the agent population. The Soviet engine lost its payer. The US engine has acquired one of a kind no prior fiscal regime hosted.
Banks frames US conditionality through the traditional fiscal and monetary policy channels: sustained inflation, debt-ceiling theater, monetary accommodation of fiscal demands, capital controls. "Every other holder of dollars is running the same calculus in parallel." That model held when foreign sovereigns were the marginal Treasury holders. They aren't anymore. The relevant marginal holder is the stablecoin issuer's reserve manager, running compliance calculus, not geopolitical calculus.
The conditionality runs through stack-provider regulatory posture. Whether Anthropic, OpenAI, Google, and successor stacks continue to operate inside US jurisdiction and denominate agent transactions in USD-pegged stablecoins. Whether GENIUS Act enforcement holds. Whether issuer cohorts consolidate or fragment. Per Sovereign Competition, this is monetary engines competing for populations rather than competing on technical properties — the contemporary competition is over which jurisdiction hosts the stack that denominates AI-agent volume. None of these variables is in Banks's framing because Banks inherits a sovereign-creditor model that the contemporary structure has moved past.
What survives the correction is the standalone interest-crowding observation. The $970B 2025 net interest is mechanical, locked into coupons already issued, structurally rising as debt rolls into higher-rate environment. The new payer helps the demand side (allows continued debt absorption at lower yields) but does not unwind existing interest burden. Banks is right that interest is mechanically eating fiscal capacity. He is wrong that this puts the US on a Soviet structural axis. Two things at once.
Three named failure modes, none Soviet-shaped:
Personal-agent disintermediation. If agent stacks shift to local/private execution and transactions route through non-issuer-mediated rails (Monero-class, CoinJoin'd BTC via the cycling-tax mechanism, alternative L1s), the reserve-composition channel attenuates. Requires stack-provider regulatory tolerance falling AND user-side preference for unmediated transacting. Low-probability central case. Higher if regulatory capture inverts.
Issuer fragility or regulatory capture. A Tether-class event (audit failure, reserve-misrepresentation discovery, enforcement action) could remove $100B+ of Treasury demand simultaneously. Cascading if it triggers stablecoin-class confidence erosion. Higher-probability than disintermediation, shorter timeline.
Pricing-stack flip. Per The Pricing of Everything, intelligence saturates one layer at a time, and pricing-of-everything denominates in whatever stack runs the layer. If pricing-of-everything denominates in CNY-stablecoins (PBOC-permissioned digital yuan plus CN-stack agent integration) or in BTC as cross-bloc agentic bridge per The Payer Question's mechanism, the demand engine for USD migrates. This is the deepest contemporary risk vector. It operates on 15-30 year horizons, not Banks's 5-10 year hints.
None of these operates through currency-conversion crisis or apparatchik preference cascade. The contemporary failure mode is denomination migration, and the binding variable is which stack denominates AI-agent volume.
A fiscal engine's binding variable is its payer-class, not its obligations. Banks treats obligations as the variable (inelastic, hard to cut) and payer-base as the assumption (creditor confidence dependent on policy posture). The contemporary US case inverts this. Obligations are growing inelastically, but the payer-class is acquiring novel members faster than the obligations grow, because the AI-agent population denominates in USD-stablecoin reserves by stack-provider default.
The Soviet engine lost a payer (oil rents) and could not substitute. The US engine acquired a payer (issuer-reserve composition driven by agent volume) at a rate that absorbs the obligation growth. The mechanism Banks named is real, the historical instances are real, and the contemporary US is not on that mechanism's axis.
What is on Banks's axis: countries with foreign-currency debt and atrophied trade, contracting payer bases without stack-denomination access. The US is in a cohort of one. The engine that hosts the global pricing-stack at the moment intelligence saturates the layers. Exorbitant privilege denominates at finer granularity each year, not coarser, because per The Pricing of Everything the layers being priced into existence are layers the engine's host stack already runs.
The conditionality is real. It does not route through fiscal exhaustion. It routes through which stack denominates the layers, and that is the axis the next decade's policy choices will trace.